Upgrade to Pro — share decks privately, control downloads, hide ads and more …

生成系AIの実応用に向けて

 生成系AIの実応用に向けて

LINE株式会社 Data Scienceセンター AI Dev室 室長 井尻善久

※画像センシング展2023
イメージセンシングセミナー:特別招待講演<ジェネレーティブAI・応用課題>
での発表資料です
https://www.adcom-media.co.jp/seminar/2023SS/session/I-371.html

LINE Developers

July 07, 2023
Tweet

More Decks by LINE Developers

Other Decks in Technology

Transcript

  1. :PTIJIJTB*+*3*
    ੜ੒"*ͷ࣮༻ʹ޲͚ͯ
    "*EFW$7- EJSFDUPS
    -*/&$PSQPSBUJPO

    View full-size slide

  2. :PTIJIJTB*KJSJ 1I%
    -*/&גࣜձࣾ σʔλαΠΤϯεηϯλʔ
    "*%FWࣨ ࣨ௕ɺ$7-Ϛωʔδϟʔ
    > ઐ໳ɿίϯϐϡʔλϏδϣϯɾϩϘςΟΫεɺͦΕΒΛࢧ͑Δػցֶश
    > झຯɿ
    > 0VUEPPSొࢁɾεΩʔɾୌ८ΓɾࣸਅࡱӨɾόΠΫτϥΠΞϧɾɾɾ
    > *OEPPSϐΞϊԋ૗ɾྺ࢙ɾᗉ੡ɾίʔώʔᖿઝɾञΛᅂΉ
    > ೥ΦϜϩϯೖࣾ
    > إͷݕग़ೝࣝͷσδΧϝɾܞଳి࿩ɺ؂ࢹΧϝϥԠ༻
    > ෺ମݕग़ɾŤŞƄŸƃũŖŢŔƃɾ0$3ͷ'"޲͚঎඼Խ
    > ͠ͳ΍͔ͳ੍ޚΛ࣮ݱ͢Δࣗ཯ιϑτϩϘοτݚڀਪਐ
    > Ϧαʔνϕϯνϟʔ্ཱͪ͛ 0.30/4*/*$9

    > ೥-*/&ೖࣾ
    > $PNQVUFS7JTJPO-BCͷ্ཱͪ͛ɺ"*։ൃࣨͷ૊৫Խ

    View full-size slide

  3. ຊ೔ͷߨԋͷ಺༰
    • ੜ੒"*ͱͦͷՄೳੑ
    • -*/&Ͱͷ"*
    • ͜Ε͔Βͷ"*ͱͦΕʹΑΓݟ͑ͯ͘ΔՄೳੑ

    View full-size slide

  4. ຊ೔ͷߨԋͷ಺༰
    • ੜ੒"*ͱͦͷՄೳੑ
    • -*/&Ͱͷ"*
    • ͜Ε͔Βͷ"*ͱͦΕʹΑΓݟ͑ͯ͘ΔՄೳੑ

    View full-size slide

  5. ʮਪ࿦ʯͷࣗಈԽͷਐల

    ػցֶशϕʔε
    `T

    ϧʔϧϕʔε
    ؍࡯΍஌ݟΛݩʹ
    ౷ܭϞσϧߏங
    σʔλ͔Β
    ༧ଌϞσϧΛ௚઀ֶश
    ը૾ɾݴޠͳͲ
    ଟ࣍ݩσʔλΛѻ͍ͮΒ͘௿ਫ਼౓
    ղऍੑɾઆ໌ੑ͕௿͘
    ܭࢉ͕๲େ
    ղऍੑɾઆ໌ੑ͕ߴ͘
    ܭࢉ΋গͳΊ
    ը૾ɾݴޠͳͲ
    ଟ࣍ݩσʔλ͕ѻ͑ߴਫ਼౓
    *GUIFOϧʔϧ
    ϕΠζ౷ܭ
    χϡʔϥϧωοτϫʔΫ
    ಛ௃
    ܽ఺
    ར఺
    ୅දख๏
    WT

    View full-size slide

  6. "*ֵ໋ɿػց͕ʮਪ࿦ʯͰਓΛ௒͑Δ࣌୅
    $IBU(15
    ήʔϜʢғޟʣͷੈքͰ
    ਓΛ௒͑ͨʂ
    ʮݴޠʯʮࢥߟʯͳͲʮੜ੒ʯͰ
    ਓΛ௒͑ͭͭ͋Δʁ
    ๲େͳعේσʔλͰֶश 8FC্ͷ๲େͳ஌ࣝ
    ίϯϐϡʔλಉ࢜Ͱରઓ͠
    ࣗ཯ֶश
    ਓͷϑΟʔυόοΫͰֶश
    "MQIB(P ܭࢉػਐԽ

    View full-size slide

  7. ͩΕ΋ѻ͑Δʮݴޠʯʮը૾ʯΛೖྗ͠
    ͩΕ΋͕Θ͔Δʮݴޠʯ΍ʮը૾ʯɾʮԻ੠ʯͱ͍͏ܗͰग़ྗͰ͖Δ "*
    ੜ੒"*ͱ͸ʁ
    7
    "*
    ਓ͕͍ͯ͠Δ࡞ۀʹ
    ͍͍ۙͮͯΔ

    View full-size slide

  8. ͩΕ΋ѻ͑Δʮݴޠʯʮը૾ʯΛೖྗ͠
    ೝࣝ݁Ռɾ൑ఆ݁ՌΛग़ྗ
    ैདྷͷ"*
    8
    "*
    ਓ͕͍ͯ͠Δ࡞ۀʹ
    ͍͍ۙͮͯΔ
    :FT/P.BZCF
    ΫϥεϥϕϧͳͲ
    جຊతʹ࣍ݩͷग़ྗ

    View full-size slide

  9. ドメインを選ばない⼤量データで学習させた多⽬的なモデル
    ⽣成AIを可能とする基盤モデル
    υ
    ϝ
    Π
    ϯ
    λεΫ
    λεΫಛԽܕ"*
    λεΫಛԽܕ"*
    ج൫Ϟσϧ

    View full-size slide

  10. ج൫Ϟσϧʢ'PVOEBUJPONPEFMʣʹΑΔ൚༻Խ
    XFC
    ,OPX
    MFEHF
    CBTF
    1VSDIBTF
    SFDPSE
    1VSDIBTF
    SFDPSE
    5SBOT
    BDUJPO
    SFDPSE
    4QFFDI EPDT
    "E
    CBOOFS
    4UJDLFS
    4FBSDI 2"
    4IPQQJOH
    FYQFSJFODF
    DIBUCPU
    $POW
    X
    TUJDLFST
    4$. %JBMPHVF 0$3
    "E
    PQUJNJ[F
    ʜ
    ʜ
    4FBSDI
    NPEFM
    2"
    NPEFM
    3FDPN
    .PEFM
    %JBMPHVF
    .PEFM
    %FNBOE
    1SFE
    NPEFM
    4QFFDI
    3FDPH
    NPEFM
    0$3
    NPEFM
    &GGFDU
    1SFE
    NPEFM
    4UJDLFS
    3FDPN
    NPEFM
    ʜ
    ϚϧνυϝΠϯσʔλ
    4FBSDI 2" ʜ 0$3
    "E
    PQUJNJ[F
    ʜ
    0OFNPEFM
    'PVOEBUJPONPEFM



    ޙ

    λ
    ε
    Ϋ


    ैདྷ
    λεΫຖʹϞσϧߏஙɺܾΊͨλεΫͷΈʹར༻
    ج൫Ϟσϧ
    υϝΠϯλεΫʹґଘͤͣ࢖͑Δ
    ͭͷٕज़ֵ৽

    View full-size slide

  11. ୯७ʹ"UUFOUJPO''/ϨΠϠʔΛॏͶΔ͚ͩͰਫ਼౓্͕͕Δʂ
    ͦΕ·Ͱ͸3//΍-45.ͳͲͰ૚͝ͱʹ޻෉͕ඞཁ
    ̍ͭ໨ͷٕज़ֵ৽ɿߏ଄ͷ୯७Խʢ5SBOTGPSNFSʢʣʣ
    Ϟσϧߏ଄ΤϯδχΞϦϯάͰͷ
    ࠩҟԽͷऴᖼʁ
    ͨͩ͠ɺ·ͩγϯάϧλεΫϞσϧ
    ·ͩ·ͩ൚༻Խ͕ඞཁͳঢ়گͩͬͨ
    ൚༻Խʹ޲͚ͯͷ՝୊͸
    ֶश͢Δσʔλྔ
    ʴ
    ͦͷͨΊͷΞϊςʔγϣϯ

    View full-size slide

  12. ̎ͭ໨ͷٕज़ֵ৽ɿڭࢣͳ͠ࣄલֶश #&35 (15


    Inclusive multi-modal data
    Search QA … OCR
    Ad
    optimize

    Pre-trained
    model
    Small data
    Fine
    tuning
    GT
    XFC
    ,OPX
    MFEHF
    CBTF
    1VSDIBTF
    SFDPSE
    1VSDIBTF
    SFDPSE
    5SBOT
    BDUJPO
    SFDPSE
    4QFFDI EPDT
    "E
    CBOOFS
    4UJDLFS
    4FBSDI 2"
    4IPQQJOH
    FYQFSJFODF
    DIBUCPU
    $POW
    X
    TUJDLFST
    4$. %JBMPHVF 0$3
    "E
    PQUJNJ[F
    ʜ
    ʜ
    4FBSDI
    NPEFM
    2"
    NPEFM
    3FDPN
    .PEFM
    %JBMPHVF
    .PEFM
    %FNBOE
    1SFE
    NPEFM
    4QFFDI
    3FDPH
    NPEFM
    0$3
    NPEFM
    &GGFDU
    1SFE
    NPEFM
    4UJDLFS
    3FDPN
    NPEFM
    ʜ


    ޙ

    λ
    ε
    Ϋ


    ैདྷ
    ैདྷ͸λεΫ͝ͱʹೖྗͱਖ਼ղ͕ඞཁ
    ڭࢣͳ͠ࣄલֶश
    ਖ਼ղ͕ཁΒͳֶ͍शํ๏

    View full-size slide

  13. ݀ຒΊ໰୊ͱ઀ଓ໰୊Λɺʢ΄΅ʣແݶʹ࡞Γग़͠ɺֶशͤ͞Δʢࣗݾڭࢣֶशʣ
    #&35ͰఏҊ͞Εͨɺڭࢣͳ͠ࣄલֶशͷΠϝʔδ
    ੨ۭจݿʮۜՏమಓͷ໷ʯΑΓ

    View full-size slide

  14. ハハ
    積もれば
    価格

    ⼭となる
    塵も積もれば
    ⽂脈 次に来る⾔葉
    ?
    ⾔語モデルとは、与えられた⽂章の次に来る⾔葉を当てるよう学習させたAIモデル
    その性能を最⼤限引き出すため、モデルサイズ(パラメータ)を10億以上に拡⼤
    #&35ͰఏҊ͞Εͨɺڭࢣͳ͠ࣄલֶशͷΠϝʔδ
    ͜ΕΛݴޠੜ੒ʹԠ༻͢Δͱɾɾɾʢ(15ʣ

    View full-size slide

  15. 4DBMJOH-BXT<,BQMBO BS9JW >ʹΑΕ͹ɺσʔληοτ΍ύϥϝʔλΛେ͖͘͢Ε͹ͲΜͲΜਫ਼౓͕ྑ͘ͳΔʂ
    εέʔϦϯά๏ଇ
    Kaplan+, Scaling Laws for Neural Language Models, arXiv 2020より抜粋

    View full-size slide

  16. େن໛ݴޠϞσϧͷ։ൃڝ૪
    ։ൃݩ ΞϧΰϦζϜ ύϥϝʔλ਺ ݴޠ
    0QFO"* (15 3BEGPSE
    . &OHMJTI
    (15 3BEGPSE
    # &OHMJTI
    (15 #SPXO
    # &OHMJTI
    (PPHMF #&35 %FWMJO
    # &OHMJTI NBOZ+1WBSJBOUTBWBJMBCMF

    5 3BGGFM 9VF
    # &OHMJTI
    4XJUDI5SBOTGPSNFS 'FEVT
    # MBOHVBHFT
    1B-.

    .JDSPTPGU .5/-(
    # &OHMJTI
    -*/& )ZQFS$-07" ,JN
    d# # +BQBOFTF
    /"7&3 )ZQFS$-07" ,JN
    d# # ,PSFBO
    .FUB 015 # &OHMJTI

    View full-size slide

  17. ̎ͭ໨ͷٕज़ֵ৽ɿڭࢣͳ͠ࣄલֶश #&35 (15


    Inclusive multi-modal data
    Search QA … OCR
    Ad
    optimize

    Pre-trained
    model
    Small data
    Fine
    tuning
    GT
    ڭࢣͳ͠ࣄલֶश
    ਖ਼ղ͕ཁΒͳֶ͍शํ๏
    λεΫదԠͷͨΊͷ
    ϑΝΠϯνϡʔχϯά͕ඞཁ

    View full-size slide

  18. େن໛ͳࣄલֶशϞσϧʹλεΫࢦࣔ͢Δ͜ͱͰɺଟ͘ͷλεΫΛ௥Ճֶशͳ࣮͘ݱͰ͖Δ͜ͱ͕ൃݟ͞Εͨʂ
    ̏ͭ໨ͷٕज़ֵ৽ɿϓϩϯϓςΟϯά (15




    ޙ

    λ
    ε
    Ϋ


    *ODMVTJWFNVMUJNPEBMEBUB
    4FBSDI 2" ʜ 0$3
    "E
    PQUJNJ[F
    ʜ
    1SFUSBJOFE
    'PVOEBUJPONPEFM
    'FXTIPUQSPNQUGPSUBTLTQFDJGJDBUJPO
    *ODMVTJWF
    NVMUJNPEBMEBUB
    4FBSDI 2" ʜ 0$3
    "E
    PQUJNJ[F
    ʜ
    1SFUSBJOFE
    NPEFM
    4NBMMEBUB
    'JOF
    UVOJOH
    (5
    ैདྷ
    λεΫ͝ͱͷϑΝΠϯνϡʔχϯά͕ඞཁ
    ϓϩϯϓςΟϯά
    λεΫࢦࣔͰԿͰ΋Ͱ͖Δʂ

    View full-size slide

  19. ϓϩϯϓςΟϯάͷ6*
    ղઆ͔Βആ۟Λੜ੒͠·͢ɻ
    */͕֝ݹ͍஑ʹඈͼࠐΜͩ࣌ͷԻͷ༷ࢠΛӵΜͩ۟Ͱ͢ɻ͕֝஑ʹඈͼࠐΉԻΛදݱͨ͠୯७ͳ۟
    Ͱ͸͋Γ·͕͢ɺपғͷ੩ऐ΍ऐΕͨݹ஑ͷ༷ࢠɺ͕֝஑ʹඈͼࠐΉੜͷ༂ಈͷΑ͏ͳ৘ܠ͕·͟
    ·͟ͱ఻Θͬͯ͘Δɺझͷ͋Δ۟ͱͳ͍ͬͯ·͢ɻ͜ͷ۟ͷقޠ͸ʰ֝ʱͰɺ͜Ε͸य़Λදݱͯ͠
    ͍ΔقޠͰ͢ɻ
    065ݹ஑΍֝ඈ͜Ήਫͷ͓ͱ
    */ࢁܗݝʹ͋ΔཱੴࣉʢΓͬ͠Ό͘͡ʣͰηϛ͕໐͍͍ͯΔ༷ࢠΛӵ·Εͨ۟Ͱ͢ɻ͜ͷ۟͸໌֬
    ͳ۟༁͸Θ͔͍ͬͯͳ͍ͷͰɺগ͠ϛεςϦΞεͰӵΈղ͘͜ͱʹϩϚϯ͕͋Δആ۟Ͱ͢ɻʮؓ
    ͞ʯͱʮઊͷ੠ʯͱ͍͏Ұݟໃ६͢Δදݱͷҙਤ͕ಛʹߟ͑ͤ͞ΒΕ·͢Ͷɻ
    065ؓ͞΍ؠʹ͠Έೖઊͷ੠
    */݄຤ͷॵ͍೔ͷޕޙͷձٞʹͱͯ΋૖େͳγεςϜ։ൃͷ࿩Λ͠Α͏ͱ͍ͯ͠Δ༷ࢠΛӵ·Ε
    ͨ۟Ͱ͢ɻ୯७ͳ۟Ͱ͸͋Γ·͕͢ɺαϥϦʔϚϯͳΤϯδχΞ͕༷ʑͳݒ೦͕͋ΔதͰɺͦΕͰ
    ΋͜ͷٕज़Λܗʹ͢Δ͜ͱͰਓ޻஌ೳٕज़ͷະདྷΛ։͖͍ͨͱ͍͏ر๬Λ͜Ίͨؾ࣋ͪΛදݱͯ͠
    ͍Δɻ
    065
    λεΫ
    એݴ

    View full-size slide

  20. ϓϩϯϓςΟϯάͷ6*
    ղઆ͔Βആ۟Λੜ੒͠·͢ɻ
    */͕֝ݹ͍஑ʹඈͼࠐΜͩ࣌ͷԻͷ༷ࢠΛӵΜͩ۟Ͱ͢ɻ͕֝஑ʹඈͼࠐΉԻΛදݱͨ͠୯७ͳ۟
    Ͱ͸͋Γ·͕͢ɺपғͷ੩ऐ΍ऐΕͨݹ஑ͷ༷ࢠɺ͕֝஑ʹඈͼࠐΉੜͷ༂ಈͷΑ͏ͳ৘ܠ͕·͟
    ·͟ͱ఻Θͬͯ͘Δɺझͷ͋Δ۟ͱͳ͍ͬͯ·͢ɻ͜ͷ۟ͷقޠ͸ʰ֝ʱͰɺ͜Ε͸य़Λදݱͯ͠
    ͍ΔقޠͰ͢ɻ
    065ݹ஑΍֝ඈ͜Ήਫͷ͓ͱ
    */ࢁܗݝʹ͋ΔཱੴࣉʢΓͬ͠Ό͘͡ʣͰηϛ͕໐͍͍ͯΔ༷ࢠΛӵ·Εͨ۟Ͱ͢ɻ͜ͷ۟͸໌֬
    ͳ۟༁͸Θ͔͍ͬͯͳ͍ͷͰɺগ͠ϛεςϦΞεͰӵΈղ͘͜ͱʹϩϚϯ͕͋Δആ۟Ͱ͢ɻʮؓ
    ͞ʯͱʮઊͷ੠ʯͱ͍͏Ұݟໃ६͢Δදݱͷҙਤ͕ಛʹߟ͑ͤ͞ΒΕ·͢Ͷɻ
    065ؓ͞΍ؠʹ͠Έೖઊͷ੠
    */݄຤ͷॵ͍೔ͷޕޙͷձٞʹͱͯ΋૖େͳγεςϜ։ൃͷ࿩Λ͠Α͏ͱ͍ͯ͠Δ༷ࢠΛӵ·Ε
    ͨ۟Ͱ͢ɻ୯७ͳ۟Ͱ͸͋Γ·͕͢ɺαϥϦʔϚϯͳΤϯδχΞ͕༷ʑͳݒ೦͕͋ΔதͰɺͦΕͰ
    ΋͜ͷٕज़Λܗʹ͢Δ͜ͱͰਓ޻஌ೳٕज़ͷະདྷΛ։͖͍ͨͱ͍͏ر๬Λ͜Ίͨؾ࣋ͪΛදݱͯ͠
    ͍Δɻ
    065

    View full-size slide

  21. ϓϩϯϓςΟϯάͷ6*
    ղઆ͔Βആ۟Λੜ੒͠·͢ɻ
    */͕֝ݹ͍஑ʹඈͼࠐΜͩ࣌ͷԻͷ༷ࢠΛӵΜͩ۟Ͱ͢ɻ͕֝஑ʹඈͼࠐΉԻΛදݱͨ͠୯७ͳ۟
    Ͱ͸͋Γ·͕͢ɺपғͷ੩ऐ΍ऐΕͨݹ஑ͷ༷ࢠɺ͕֝஑ʹඈͼࠐΉੜͷ༂ಈͷΑ͏ͳ৘ܠ͕·͟
    ·͟ͱ఻Θͬͯ͘Δɺझͷ͋Δ۟ͱͳ͍ͬͯ·͢ɻ͜ͷ۟ͷقޠ͸ʰ֝ʱͰɺ͜Ε͸य़Λදݱͯ͠
    ͍ΔقޠͰ͢ɻ
    065ݹ஑΍֝ඈ͜Ήਫͷ͓ͱ
    */ࢁܗݝʹ͋ΔཱੴࣉʢΓͬ͠Ό͘͡ʣͰηϛ͕໐͍͍ͯΔ༷ࢠΛӵ·Εͨ۟Ͱ͢ɻ͜ͷ۟͸໌֬
    ͳ۟༁͸Θ͔͍ͬͯͳ͍ͷͰɺগ͠ϛεςϦΞεͰӵΈղ͘͜ͱʹϩϚϯ͕͋Δആ۟Ͱ͢ɻʮؓ
    ͞ʯͱʮઊͷ੠ʯͱ͍͏Ұݟໃ६͢Δදݱͷҙਤ͕ಛʹߟ͑ͤ͞ΒΕ·͢Ͷɻ
    065ؓ͞΍ؠʹ͠Έೖઊͷ੠
    */݄຤ͷॵ͍೔ͷޕޙͷձٞʹͱͯ΋૖େͳγεςϜ։ൃͷ࿩Λ͠Α͏ͱ͍ͯ͠Δ༷ࢠΛӵ·Ε
    ͨ۟Ͱ͢ɻ୯७ͳ۟Ͱ͸͋Γ·͕͢ɺαϥϦʔϚϯͳΤϯδχΞ͕༷ʑͳݒ೦͕͋ΔதͰɺͦΕͰ
    ΋͜ͷٕज़Λܗʹ͢Δ͜ͱͰਓ޻஌ೳٕज़ͷະདྷΛ։͖͍ͨͱ͍͏ر๬Λ͜Ίͨؾ࣋ͪΛදݱͯ͠
    ͍Δɻ
    065

    View full-size slide

  22. ϓϩϯϓςΟϯάͷ6*
    ղઆ͔Βആ۟Λੜ੒͠·͢ɻ
    */͕֝ݹ͍஑ʹඈͼࠐΜͩ࣌ͷԻͷ༷ࢠΛӵΜͩ۟Ͱ͢ɻ͕֝஑ʹඈͼࠐΉԻΛදݱͨ͠୯७ͳ۟
    Ͱ͸͋Γ·͕͢ɺपғͷ੩ऐ΍ऐΕͨݹ஑ͷ༷ࢠɺ͕֝஑ʹඈͼࠐΉੜͷ༂ಈͷΑ͏ͳ৘ܠ͕·͟
    ·͟ͱ఻Θͬͯ͘Δɺझͷ͋Δ۟ͱͳ͍ͬͯ·͢ɻ͜ͷ۟ͷقޠ͸ʰ֝ʱͰɺ͜Ε͸य़Λදݱͯ͠
    ͍ΔقޠͰ͢ɻ
    065ݹ஑΍֝ඈ͜Ήਫͷ͓ͱ
    */ࢁܗݝʹ͋ΔཱੴࣉʢΓͬ͠Ό͘͡ʣͰηϛ͕໐͍͍ͯΔ༷ࢠΛӵ·Εͨ۟Ͱ͢ɻ͜ͷ۟͸໌֬
    ͳ۟༁͸Θ͔͍ͬͯͳ͍ͷͰɺগ͠ϛεςϦΞεͰӵΈղ͘͜ͱʹϩϚϯ͕͋Δആ۟Ͱ͢ɻʮؓ
    ͞ʯͱʮઊͷ੠ʯͱ͍͏Ұݟໃ६͢Δදݱͷҙਤ͕ಛʹߟ͑ͤ͞ΒΕ·͢Ͷɻ
    065ؓ͞΍ؠʹ͠Έೖઊͷ੠
    */݄຤ͷॵ͍೔ͷޕޙͷձٞʹͱͯ΋૖େͳγεςϜ։ൃͷ࿩Λ͠Α͏ͱ͍ͯ͠Δ༷ࢠΛӵ·Ε
    ͨ۟Ͱ͢ɻ୯७ͳ۟Ͱ͸͋Γ·͕͢ɺαϥϦʔϚϯͳΤϯδχΞ͕༷ʑͳݒ೦͕͋ΔதͰɺͦΕͰ
    ΋͜ͷٕज़Λܗʹ͢Δ͜ͱͰਓ޻஌ೳٕज़ͷະདྷΛ։͖͍ͨͱ͍͏ر๬Λ͜Ίͨؾ࣋ͪΛදݱͯ͠
    ͍Δɻ
    065
    ຊ୊

    View full-size slide

  23. ϓϩϯϓςΟϯάͷ6*
    ղઆ͔Βആ۟Λੜ੒͠·͢ɻ
    */͕֝ݹ͍஑ʹඈͼࠐΜͩ࣌ͷԻͷ༷ࢠΛӵΜͩ۟Ͱ͢ɻ͕֝஑ʹඈͼࠐΉԻΛදݱͨ͠୯७ͳ۟
    Ͱ͸͋Γ·͕͢ɺपғͷ੩ऐ΍ऐΕͨݹ஑ͷ༷ࢠɺ͕֝஑ʹඈͼࠐΉੜͷ༂ಈͷΑ͏ͳ৘ܠ͕·͟
    ·͟ͱ఻Θͬͯ͘Δɺझͷ͋Δ۟ͱͳ͍ͬͯ·͢ɻ͜ͷ۟ͷقޠ͸ʰ֝ʱͰɺ͜Ε͸य़Λදݱͯ͠
    ͍ΔقޠͰ͢ɻ
    065ݹ஑΍֝ඈ͜Ήਫͷ͓ͱ
    */ࢁܗݝʹ͋ΔཱੴࣉʢΓͬ͠Ό͘͡ʣͰηϛ͕໐͍͍ͯΔ༷ࢠΛӵ·Εͨ۟Ͱ͢ɻ͜ͷ۟͸໌֬
    ͳ۟༁͸Θ͔͍ͬͯͳ͍ͷͰɺগ͠ϛεςϦΞεͰӵΈղ͘͜ͱʹϩϚϯ͕͋Δആ۟Ͱ͢ɻʮؓ
    ͞ʯͱʮઊͷ੠ʯͱ͍͏Ұݟໃ६͢Δදݱͷҙਤ͕ಛʹߟ͑ͤ͞ΒΕ·͢Ͷɻ
    065ؓ͞΍ؠʹ͠Έೖઊͷ੠
    */݄຤ͷॵ͍೔ͷޕޙͷձٞʹͱͯ΋૖େͳγεςϜ։ൃͷ࿩Λ͠Α͏ͱ͍ͯ͠Δ༷ࢠΛӵ·Ε
    ͨ۟Ͱ͢ɻ୯७ͳ۟Ͱ͸͋Γ·͕͢ɺαϥϦʔϚϯͳΤϯδχΞ͕༷ʑͳݒ೦͕͋ΔதͰɺͦΕͰ
    ΋͜ͷٕज़Λܗʹ͢Δ͜ͱͰਓ޻஌ೳٕज़ͷະདྷΛ։͖͍ͨͱ͍͏ر๬Λ͜Ίͨؾ࣋ͪΛදݱͯ͠
    ͍Δɻ
    065ീ݄ͷ೤͍࠭ͷத΁खΛೖΕΔ

    View full-size slide

  24. ϓϩϯϓςΟϯάͷ6*
    ղઆ͔Βആ۟Λੜ੒͠·͢ɻ
    */͕֝ݹ͍஑ʹඈͼࠐΜͩ࣌ͷԻͷ༷ࢠΛӵΜͩ۟Ͱ͢ɻ͕֝஑ʹඈͼࠐΉԻΛදݱͨ͠୯७ͳ۟
    Ͱ͸͋Γ·͕͢ɺपғͷ੩ऐ΍ऐΕͨݹ஑ͷ༷ࢠɺ͕֝஑ʹඈͼࠐΉੜͷ༂ಈͷΑ͏ͳ৘ܠ͕·͟
    ·͟ͱ఻Θͬͯ͘Δɺझͷ͋Δ۟ͱͳ͍ͬͯ·͢ɻ͜ͷ۟ͷقޠ͸ʰ֝ʱͰɺ͜Ε͸य़Λදݱͯ͠
    ͍ΔقޠͰ͢ɻ
    065ݹ஑΍֝ඈ͜Ήਫͷ͓ͱ
    */ࢁܗݝʹ͋ΔཱੴࣉʢΓͬ͠Ό͘͡ʣͰηϛ͕໐͍͍ͯΔ༷ࢠΛӵ·Εͨ۟Ͱ͢ɻ͜ͷ۟͸໌֬
    ͳ۟༁͸Θ͔͍ͬͯͳ͍ͷͰɺগ͠ϛεςϦΞεͰӵΈղ͘͜ͱʹϩϚϯ͕͋Δആ۟Ͱ͢ɻʮؓ
    ͞ʯͱʮઊͷ੠ʯͱ͍͏Ұݟໃ६͢Δදݱͷҙਤ͕ಛʹߟ͑ͤ͞ΒΕ·͢Ͷɻ
    065ؓ͞΍ؠʹ͠Έೖઊͷ੠
    */݄຤ͷॵ͍೔ͷޕޙͷձٞʹͱͯ΋૖େͳγεςϜ։ൃͷ࿩Λ͠Α͏ͱ͍ͯ͠Δ༷ࢠΛӵ·Ε
    ͨ۟Ͱ͢ɻ୯७ͳ۟Ͱ͸͋Γ·͕͢ɺαϥϦʔϚϯͳΤϯδχΞ͕༷ʑͳݒ೦͕͋ΔதͰɺͦΕͰ
    ΋͜ͷٕज़Λܗʹ͢Δ͜ͱͰਓ޻஌ೳٕज़ͷະདྷΛ։͖͍ͨͱ͍͏ر๬Λ͜Ίͨؾ࣋ͪΛදݱͯ͠
    ͍Δɻ
    065༦ম΍ϓϩδΣΫτԌ্ͤΓ

    View full-size slide

  25. ϓϩϯϓςΟϯάʹΑΔ঎඼આ໌จͷੜ੒

    View full-size slide

  26. %#ͷεέʔϧ֦ு΋ॏཁ͕ͩɺਓͷհࡏ΋ॏཁͰ͋Δ͜ͱ͕ঃʑʹ൑໌
    "*Λ͞ΒʹҭͯΔʂ
    26
    5-%3EBUBTFU

    SFEEJUDPNΑΓసࡌ
    .BTLFEMBOHVBHFNPEFMMJOH

    View full-size slide

  27. ࣌୅ʹٯߦ͢ΔΑ͏͕ͩɺڭࢣ͋ΓֶशΛ࢖ͬͯϑΝΠϯνϡʔχϯά͢Δ
    "*ΛҭͯΔ

    'JOFUVOFEMBOHVBHFNPEFMTBSF[FSPTIPUMFBSOFST<8FJ *$-3>ΑΓసࡌ

    View full-size slide

  28. Ұ෦ͷύϥϝʔλͷΈϑΝΠϯνϡʔχϯά
    "*ΛҭͯΔ
    28
    -P3"5SBOTGPSNFSͷ֤ϨΠϠʹ͓͍ͯɺ"EBQUPSNPEVMF
    ΛCZQBTTͤ͞ɺͦΕΛUVOJOH͢Δ͜ͱͰগ਺σʔλͰ΋աద
    ߹ͳ͘ɺ͔ͭޮ཰Α͘దԠͤ͞ΒΕΔख๏
    )V -03"-083"/,"%"15"5*0/0'-"3(&-"/(6"(& .0%&-4 *$-3 ΑΓసࡌ

    View full-size slide

  29. ਓͷհࡏΛ͏·͘׆༻Ͱ͖ΔֶशΞϧΰϦζϜͱͯ͠ͷ3-
    "*ΛҭͯΔ
    • ใुؔ਺Λਓͷ؆୯ͳϑΟʔυόοΫʹ
    ج͍ͮͯϑΟοςΟϯά
    • ݁Ռతʹେྔͷσʔλʹεέʔϧͤ͞Δ
    ͜ͱ͕Մೳ
    • ͜ΕʹΑΓෳࡶͳใुؔ਺΋දݱͰ͖Δ
    • ใु੒ܗʢSFXBSETIBQJOHʣͷҰͭͱߟ
    ͑ΒΕ͍ͯΔɻ
    • ౰࣌͸ɺ"$ 5310ͳͲͷڧԽֶशΛ༻
    ͍ͯɺϩϘςΟΫεͱήʔϜͷλεΫʹ
    ׆༻

    %FFQSFJOGPSDFNFOUMFBSOJOHGSPNIVNBOQSFGFSFODFT<$ISJTUJBOP BS9JW>ΑΓసࡌ

    View full-size slide

  30. *OTUSVDU(15 <0VZBO .BS>
    5SBJOJOHMBOHVBHFNPEFMTUPGPMMPXJOTUSVDUJPOTXJUIIVNBOGFFECBDL<0VZBOH >ΑΓసࡌ

    View full-size slide

  31. ͦΕͧΕͷεςοϓʹඞཁͳσʔλྔ
    *OTUSVDU(15 <0VZBO .BS>
    31
    4VQFSWJTFEGJOFUVOJOH 3FXBSENPEFM
    3-XJUI
    )VNBO'FFECBDL
    ,QSPNQUT
    IVNBOMBCFMFS

    ,QSPNQUT
    IVNBOMBCFMFS

    ,QSPNQUT

    View full-size slide

  32. "#ςετܗࣜͰධՁͤ͞Δ͜ͱͰɺ3-)'༻ͷ৽ͨͳֶशσʔλΛूΊΔ
    )VNBOBOOPUBUJPOͷ6*
    32

    View full-size slide

  33. $IBU(15
    ਓͷ໛ൣճ౴Λֶश͠ɺैདྷΑΓ΋๬·͍͠ճ౴͕Ͱ͖Δ
    33
    (15
    検索 質問回答 … 推薦 対話 …
    Foundation model
    ؒҧͬͨ͜ͱΛ
    ಊʑͱ౴͑Δ
    ภݟ͕໰୊ൃݴ͕
    ؚ·ΕΔճ౴Λ͢Δ
    $IBU(15
    Tuned model
    ࣭໰ճ౴
    ճ౴ධՁ
    ਓͷ࣭໰ճ౴΍ճ౴ධՁʹج͖ͮɺϞσϧΛ͞Βʹ܇࿅
    ͢Δ͜ͱͰɺ๬·͍͠ճ౴ΛಘΒΕΔΑ͏ʹͨ͠
    ճ౴ਫ਼౓͕޲্
    ෆద੾ճ౴ͷ௿ݮ
    ݴ͍ճ͠ɾදݱͷվળ
    %FFQTQFFE .JDSPTPGU
    ͳͲͰ࣮૷ɿ
    IUUQTXXXBYJPO[POFEFFQTQFFEFYUSFNFTDBMFNPEFMUSBJOJOHGPSFWFSZPOF

    View full-size slide

  34. ࢹ֮৘ใͱݴޠ৘ใΛೖྗͱͯ͠ɺݴޠΛग़ྗɻਤදೖͷࢼݧ໰୊Ͱ΋ߴಘ఺Λ࣮ݱ
    (15 $IBU(15 QMVT

    34

    View full-size slide

  35. ΞϝϦΧͷ༷ʑͳࢼݧͰɺਓؒͷ্Ґͷ੒੷Λ࢒͍ͯ͠Δ
    "*ͷ࣮ྗɿςΩετϕʔεͰ͸طʹਓؒΛ௒͑ͭͭ͋Δ
    35






    6OJGPSN#BS&YBN
    ౷Ұ࢘๏ࢼݧ
    4"53FBEJOH8SJUJOH
    େֶਐֶదੑࢼݧ
    ϦʔσΟϯά
    ˍϥΠςΟϯά
    4"5.BUI
    େֶਐֶదੑࢼݧ
    ਺ֶ
    (3&2VBOUJUBUJWF
    େֶӃڞ௨ࢼݧ
    ਺ֶ
    (3&7FSCBM
    େֶӃڞ௨ࢼݧ
    ݴޠ
    QFSDFOUJMF

    View full-size slide

  36. ਓؒͰ͸͙͢ʹ࡞Εͳ͍จষΛॻ͚Δ
    "*ͷ࣮ྗɿςΩετϕʔεͰ͸طʹਓؒΛ௒͑ͭͭ͋Δ

    View full-size slide

  37. แׅతͳϞσϧͱͦΕʹΑΔଟ༷ͳʢແݶͷʁʣλεΫͷ࣮ݱɺ͜ΕʹΑΓਓΛ௒͑ΒΕ
    ΔՄೳੑ͕ग़͖ͯͨʂ
    ج൫Ϟσϧ͕ࣔࠦ͢Δ͜ͱ
    υ
    ϝ
    Π
    ϯ
    λεΫ
    λεΫಛԽܕ"*
    λεΫಛԽܕ"*
    ج൫Ϟσϧ
    ͜
    ͷ

    ͕
    ڭ

    ͳ
    ͠


    ֶ

    Ͱ
    Մ

    ʹ
    ϓϩϯϓςΟϯά
    ڭ

    ͳ
    ͠


    ֶ


    ݾ
    ڭ

    ֶ

    View full-size slide

  38. %"--&<3BNFTI BS9JW > $-*1<3BEGPSE BS9JW >
    ϚϧνϞʔμϧͳੜ੒
    38
    3BEGPSE -FBSOJOH5SBOTGFSBCMF7JTVBM.PEFMT'SPN/BUVSBM-BOHVBHF4VQFSWJTJPO BS9JW
    3BNFTI )JFSBSDIJDBM5FYU$POEJUJPOBM*NBHF(FOFSBUJPOXJUI$-*1-BUFOUT BS9JW ΑΓసࡌ
    画像と、テキストのエンコーディングの対から、対象学習を⾏うことで、双⽅の共通の埋め込み空間を学習

    View full-size slide

  39. ௿࣍ݩۭؒͰ %JGGVTJPONPEFMΛֶशਪ࿦͢Δ͜ͱͰɺলϦιʔε͔ͭߴ଎ͳֶशਪ࿦͕Մೳ
    4UBCMFEJGGVTJPO<3PNCBDI "QS>
    359 (#ϏσΦϝϞϦʔ
    Ҏ্ఔ౓ͷ(16͕͋Ε͹ࣗ୐Ͱ΋ಈ͔͢͜ͱ͕Ͱ͖ΔΑ͏ʹ
    ͳͬͨ
    IUUQTBSYJWPSHBCTΑΓసࡌ
    ίʔυɿ IUUQTHJUIVCDPN$PNQ7JTTUBCMFEJGGVTJPO
    IUUQTIVHHJOHGBDFDPTQBDFTTUBCJMJUZBJTUBCMFEJGGVTJPO

    View full-size slide

  40. ੜ੒͞ΕΔը૾͕ಉ͡ʹͳΔΑ͏ʹɺೖྗจࣈྻͷ&NCFEEJOHΛ࠷దԽ
    ৽ͨͳʮ࡞෩ʯΛ֮͑ͤ͞Δɿ5FYUVSBMJOWFSTJPO<(BM >
    (BM "O*NBHFJT8PSUI 0OF8PSE1FSTPOBMJ[JOH5FYUUP*NBHF(FOFSBUJPOVTJOH5FYUVBM*OWFSTJPO BS9JW ΑΓసࡌ

    View full-size slide

  41. ͨͬͨ਺ຕͷը૾Ͱաద߹͢Δ͜ͱͳ͘ϑΝΠϯνϡʔχϯάɺࣗݾੜ੒ͨ͠αϯϓϧΛ༻͍Δ͜ͱ
    Ͱաֶशͱ֓೦γϑτΛ๷ࢭ
    ৽ͨͳର৅Λ֮͑ͤ͞Δɿ%SFBN#PPUI <3VJ[ "VH>
    IUUQTBSYJWPSHBCT ΑΓసࡌ
    ίʔυɿ IUUQTHJUIVCDPNTNZESFBNCPPUIHVJ

    View full-size slide

  42. গ਺ͷը૾Ͱաద߹͢Δ͜ͱͳ͘ϑΝΠϯνϡʔχϯά
    ͞·͟·ͳը૾ϓϩϯϓτͷ׆༻ɿ$POUSPM/FU<;IBOH >
    42
    基本的な考え⽅はLoRAなどに使われるAdaptorと同
    じ考え⽅、0で初期化するZeroConvが特徴的
    Zhang+, Adding Conditional Control to Text-to-Image
    Diffusion Models, arXiv, 2023より転載

    View full-size slide

  43. 'MBNJOHP<"MBZSBD BS9JW >
    ϚϧνϞʔμϧ"*ʹ͓͚ΔϓϩϯϓςΟϯά
    "MBZSBD 'MBNJOHPB7JTVBM-BOHVBHF.PEFMGPS'FX4IPU-FBSOJOH BS9JW
    ը૾ɾςΩετͷฒͼ͔Βɺ࣍ͷฒͼΛ༧ଌ͢ΔֶशΛ
    ͓ͯ͘͜͠ͱͰ৽ͨͳೖྗʹର͢Δద੾ͳ౴͑Λਪ࿦

    View full-size slide

  44. (BUPBHFOFSBMJTUBHFOU<3FFE BS9JW >
    ͞ΒʹϚϧνϞʔμϧʹɾɾɾ
    44
    3FFE "(FOFSBMJTU"HFOU BS9JW ΑΓసࡌ
    ೚ҙͷλεΫʹ͓͚Δ؍ଌͱΞΫγϣϯͷܥྻΛਖ਼نԽ͠τʔΫϯԽͯ͠ɺݴޠͱಉ͡Α͏ʹ༩͑ɺ
    5SBOTGPSNFSߏ଄ͷϞσϧʹ͓͍ͯɺ࣍ͷτʔΫϯΛ༧ଌ͢ΔΑ͏ֶश

    View full-size slide

  45. ྫɿࣸਅͷͲ͕͓͔͍͔͜͠Λཧղ͠ݴޠԽ
    ը૾Λཧղɾղऍ͢ΔೳྗͰ΋ਓʹฒΜͩʁ
    45

    View full-size slide

  46. ྫɿྫྷଂݿͷத਎ͷࣸਅΛݩʹɺͦΕΒͰԿ͕࡞ΕΔ͔Λਪન
    ը૾Λཧղɾղऍ͢ΔೳྗͰ΋ਓʹฒΜͩʁ
    46

    View full-size slide

  47. 47
    ྫɿςΩετͷࢦࣔΛཧղͯ͠ɺෳࡶͳΠϥετΛੜ੒ -BTDPBJ

    ը૾Λඳ͘ೳྗͰ΋ਓʹฒΜͩʁ

    View full-size slide

  48. ಛఆͷϞμϦςΟ͔ΒϚϧνϞʔμϧ΁ɻͦΕʹ൐͍ͲΜͳλεΫ͕ߟ͑ΒΕΔ͔ʁ
    ͲΜͲΜͱυϝΠϯͱλεΫ͸޿͕Δɾɾ
    λεΫ
    ϓϩϯϓςΟϯά
    ڭ

    ͳ
    ͠


    ֶ


    ݾ
    ڭ

    ֶ

    υ
    ϝ
    Π
    ϯ
    ϚϧνϞʔμϧԽ
    Ϛ
    ϧ
    ν
    Ϟ

    μ
    ϧ
    Խ

    View full-size slide

  49. ൚༻ਓ޻஌ೳ "(*BSUJGJDJBMHFOFSBMJOUFMMJHFODF

    <(VCSVE >
    49
    "(*
    ਓؒϨϕϧͷ൚༻తͳ஌ೳ
    ঢ়گมԽʹରͯ͠ॊೈ
    /BSSPX"*
    ಛఆͷػೳɾঢ়گΛ૝ఆ͠
    σʔλʹ࠷దԽͨ͠ػց
    ϓϩάϥϜͨ͠Ҏ্ͷ͜ͱ͕Ͱ͖Δɻ
    ಺ૠతʹߟ͍͑ͯΔ͕ɺͦͷൣғ͕ਓஐ
    Λ௒͍͑ͯΔ
    ڭ͑ͨ͜ͱ͔͠΍Βͳ͍ɻ
    ڭ͍͑ͯͳ͍͜ͱ͸Ͱ͖ͳ͍ɻ
    ਓͷΑ͏ʹҰͭͷ಄೴Ͱঢ়گʹ߹Θ͍ͤͯΖ͍Ζͳग़ྗ͕Մೳ

    View full-size slide

  50. ੜ੒"*ͷ༻్Ձ஋͸ओʹछྨ
    50
    ޮ཰Խ ίϯςϯπੜ੒ ύʔιφϧԽ

    View full-size slide

  51. ੜ੒ܕݕࡧɺݕࡧ݁ՌΛҰͭͣͭग़͢ͷͰ͸ͳ͘ɺ·ͱΊͨϨϙʔτΛදࣔ
    ৽ͨͳݕࡧύϥμΠϜ
    51
    .JDSPTPGU#JOH &EHF

    ੜ੒ܕݕࡧ
    ޮ཰Խ

    View full-size slide

  52. $FSUBJOMZɿΧελϚʔαʔϏεͷࣗಈԽ
    ⼈⼒によるカスタマーサービスを代替
    52
    ޮ཰Խ

    View full-size slide

  53. .JDSPTPGUɿϩϘοτͷ੍ޚίʔυ࡞੒ͷࣗಈԽ
    ςΩετͰࢦࣔΛग़ͤ͹ɺੜ੒"*͕ࣗಈͰίʔυʹม׵
    53
    ޮ཰Խ

    View full-size slide

  54. -*/&ɿ޿ࠂίϐʔ࡞੒ͷࣗಈԽ
    ਓྗ࡞ۀͷࣗಈԽʹΑΓੜ࢈ੑΛେ͖͘վળ
    54
    ίϯςϯπ
    ੜ੒

    View full-size slide

  55. ը૾ʴΩϟονίϐʔʴϨΠΞ΢τͷੜ੒ʹΑΔ޿ࠂແݶੜ੒
    ి௨ɿ޿ࠂੜ੒΁ͷԠ༻

    Φ
    Ϧ
    Τ
    ϯ


    ి௨ใΑΓసࡌ
    IUUQTEFOUTVIPDPNBSUJDMFT


    ύ
    ϥ
    ϝ

    λ


    Ϩ
    Π
    Ξ
    ΢
    τ


    Ω


    ν


    ը



    ޮ
    Ռ


    ίϯςϯπ
    ੜ੒

    View full-size slide

  56. NJHOɿݐஙઃܭۀ຿ͷޮ཰Խ
    56
    キーワードや写真をインプット ⽣成AIがデザインイメージをアウトプット
    ίϯςϯπ
    ੜ੒
    “リビング“

    View full-size slide

  57. &YQFEJBɿཱྀఔܾఆΛαϙʔτ
    ཱྀߦʹؔ͢Δ࣭໰ʹ౴͑Δ͜ͱͰɺސ٬ຬ଍౓ͱച্Λ޲্
    57
    ύʔιφϧԽ

    View full-size slide

  58. ੜ੒"*ͷԠ༻Ωϟϯόε
    ΤϯλʔςΠϯϝϯ
    τ
    ϚʔέςΟϯά Πϯλʔωοτ ڭҭ ҩྍ෱ࢱ ϑΝογϣϯɾσβ
    Πϯ
    Ұൠۀ຿
    ޮ

    Խ
    ৘ใݕࡧ ը૾ɾԻָݕࡧ ৘ใਪનɺ޿ࠂ഑৴ ݕࡧɺϨϙʔςΟϯ
    άɺཱྀߦܦ࿏ɾҿ৯
    ళɾߪങఏҊ
    ڭࡐࢧԉɺ"*νϡʔ
    λʔͳͲ
    ݸਓϨϕϧͰͷॳظ
    ਍அࢧԉ
    ίʔσΟϯάࢧԉ
    ৘ใՃ޻ ฤۂɺ੾Γൈ͖ಈը
    ੜ੒ɺτϨΠϥʔੜ

    ৘ใΩϡϨʔγϣϯ จॻɾϝʔϧ౳࡞੒
    ࢧԉɾਤද࡞੒
    ίϛϡχέʔ
    γϣϯ
    ΧελϚʔαϙʔτɺ
    νϟοτϘοτ
    ձ࿩ࢧԉ จॻɾϝʔϧ౳࡞੒
    ࢧԉ
    ΧελϚʔαϙʔτ
    ί
    ϯ
    ς
    ϯ
    π


    ίϯςϯπੜ੒ өըɾ57٭ຊੜ੒ɺ
    ಈըɾָۂੜ੒ɺ࡞
    ࢺ࡞ۂࢧԉɺήʔϜ
    ΩϟϥΫλʔɾγφ
    ϦΦੜ੒ɺΞχϝੜ

    ঎඼ɾ޿ࠂจͷੜ੒ɺ
    ޿ࠂੜ੒αΠτ࡞੒ɺ
    ϒϩάɾ4/4هࣄੜ
    ੒ɺ
    αΠτੜ੒ɺϒϩ
    άɾ4/4هࣄੜ੒ɺ
    ֶशίϯςϯπੜ੒ σβΠϯҊੜ੒ɺ"*
    Ϟσϧੜ੒
    Πϥετ࡞੒ɺϏδ
    ωεઓཱུҊɾΞΠ
    σΟΤʔγϣϯࢧԉɺ
    ύ

    ι
    φ
    ϧ
    Խ
    ਪનɾ൑அࢧԉ ύʔιφϥΠζϓϨ
    ΠϦετ
    ޿ࠂ৹ࠪࣗಈԽ
    ޿ࠂޮՌ࠷దԽɿૌ
    ٻϙΠϯτɾλʔ
    ήοτ౳ೖΕΕ͹ɺ
    ഔମɾΫϦΤΠςΟ
    ϒؚΊਪન
    5SJQΞγετɿཱྀఔ
    ߏஙˍ༧໿
    ࠾఺ɾධՁࢧԉ
    ڭҭϓϩάϥϜߏ
    ஙɿڭࡐબఆɾ೉қ
    ౓બఆɾɾɾ
    ਍அࢧԉ
    ݈߁ɾάϧϝΞγε
    τɿݙཱˍ഑ୡ༧໿ɺ
    ӡಈϝχϡʔߏங
    όʔνϟϧτϥΠϯ
    Φϯɺίʔσਪન
    ཁ໿ɺςετࢧԉ
    ϨϙʔςΟϯά τϨϯυϨϙʔτ ·ͱΊهࣄੜ੒ ใࠂॻ࡞੒ࢧԉ ใࠂॻ࡞੒ࢧԉɺ࿦
    จɾಛڐࣥචࢧԉ
    ੜ੒ܕݕࡧɿݕࡧ݁
    Ռͷ·ͱΊ

    View full-size slide

  59. ຊ೔ͷߨԋͷ಺༰
    • ੜ੒"*ͱͦͷՄೳੑ
    • -*/&Ͱͷ"*
    • ͜Ε͔Βͷ"*ͱͦΕʹΑΓݟ͑ͯ͘ΔՄೳੑ

    View full-size slide

  60. -*/&ʹ͓͚Δ"*ݚڀ
    /-1
    "41
    $7-
    5SVTUXPSUIZ
    "*
    .VMUJNPEBM"*

    View full-size slide

  61. 1VCMJDBUJPOTBUUIFUPQUJFSDPOGFSFODFT
    3%BDDPNQMJTINFOUT
    ': $POGFSFODFT

    *$"441
    */5&341&&$)
    8"41""
    #JH%BUB


    *$"441
    &64*1$0
    */5&341&&$)
    %$"4&
    "14*1"
    $713
    51%1
    '-*$.-

    -%3$


    *$"441
    *$3"
    *6*
    *$%&
    *$$7
    "14*1"
    */5&341&&$)
    *&&&54141-


    *$-3
    *$"441
    */5&341&&$)
    4*(.0%
    "14*1"
    *+$7
    4-5
    7-%#

    #JH%BUB


    8"$7
    """*
    *$"441
    4-5

    View full-size slide

  62. -*/&"*ͷ3%7JTJPO
    $POTFSWBUJWF
    %JTSVQUJWF
    5JNF

    *OUFSBDUJWFWJSUVBM
    FYQFSJFODF
    "VUPOPNPVT"*
    XPSLGMPX
    %JHJUBM.F
    .F"7"5"3
    %JHJUBM*EFOUJUZ
    #FUUFS$BSF
    5SVTUXPSUIZ"*
    "*'BJSOFTT
    &YQMBJOBCMF"*
    %BSL%BUB
    0NOJQPUFOU"*
    (JHBOUJD-BOHVBHFNPEFM
    6OMBCFMFE%BUB
    %BUB.BSLFUQMBDF
    (FOFSBUJWF*OUFMMJHFODF
    /FX&EVDBUJPO
    %FQFOEBCMF455
    1SJWBDZQSFTFSWJOH
    4FBN%JTDSJNJOBUPS
    -*/&ͷ࣋ͭଟ༷ͳ"*ཁૉٕज़Λجʹɺ෼໺ԣஅతͳٕज़Λ஫ྗٕज़ͱͯ͠શମײΛ࣋ͬͯݚڀ։ൃ͍ͯ͠·͢

    View full-size slide

  63. 63
    OPUPOMZ$PHOJUJPO
    CVU
    $SFBUJWJUZTVQQPSU

    View full-size slide

  64. 世界でも⼤規模⾔語モデルの開発は限られ
    ⽇本ではHyperCLOVAが最⼤
    ˞͜͜Ͱ͸#ఔ౓Ҏ্Λେن໛ͱදݱ
    Alibaba
    Tongyi Qianwen
    Cohere.ai
    Large LM
    DeepMind
    Gopher, Chinchilla, Flamingo
    Amazon
    Amazon Titan
    Baidu
    Ernie Bot
    Google
    T5, PaLM/PaLM-E,
    LaMDA/Bard
    MS-Nvidia
    Megatron-Turing NLG
    Anthropic
    Claude
    Kakaobrain
    koGPT3, Coyo
    LINE/WMJ
    HyperCLOVA
    NAVER
    HyperCLOVA
    OpenAI
    GPT/ChatGPT
    EleutherAI
    pythia
    Meta/Stanford/UCB
    OPT/LLaMA/Alpaca/Koala
    AI21 Labs
    Jurassic-1

    View full-size slide

  65. จॻ࡞੒ࢧԉ
    ཁ݅ͷΈͷϝʔϧͰ࡞จ
    จॻߍਖ਼ɾ੔ܗɾλΠτϧੜ੒
    σδλϧϚʔέςΟϯάࢧԉ
    Ωϟονίϐʔ࡞੒ࢧԉ
    େྔͷ঎඼ʹର͢Δ঎඼આ໌
    ΧελϚέΞʔࢧԉ
    ސ٬Ұ࣍αϙʔτɾৼΓ෼͚
    ίʔϧηϯλͷ͞ΒͳΔࣗಈԽ
    ৘ใऩूࢧԉ
    େྔͷσʔλʹج࣭ͮ͘໰ճ౴
    จॻͷཁ໿ɾϨϙʔτੜ੒ݕࡧ
    ج൫Ϟσϧͷ׆༻
    ͦͷଞ
    ίʔσΟϯάࢧԉɾจ๏νΣοΫɾ຋༁ɾɾɾ

    View full-size slide

  66. -*/&ʹ͓͚Δج൫Ϟσϧߏங
    %FDPEFS

    View full-size slide

  67. ޿ࠂΩϟονίϐʔੜ੒
    ΩʔϫʔυೖྗͷΈͰɺ؆қʹΩϟονΛෳ਺ੜ੒

    View full-size slide

  68. ୈ̑ճ ର࿩γεςϜ ϥΠϒίϯϖςΟγϣϯ
    オープントラック シチュエーショントラック
    IUUQTTJUFTHPPHMFDPNWJFXETMD&#&&$ BVUIVTFS

    View full-size slide

  69. %JBMPHVF3PCPU$PNQFUJUJPO
    *304ͷXPSLTIPQͱͯ͠։࠵͞Εͨର࿩ϩϘοτνϟϨϯδ

    View full-size slide

  70. 70
    9"* 1SJWBDZDBSF
    UP
    $PMMBCPSBUJWF"*
    &UIJDBM'BJS"*

    View full-size slide

  71. ج൫Ϟσϧʹ͓͚Δμ΢ϯαΠυݒ೦

    View full-size slide

  72. "*ʹ͸ΨʔυϨʔϧ͕ඞཁ

    View full-size slide

  73. /(ϫʔυʴ༗֐දݱ൑ఆ

    View full-size slide

  74. )BMMVDJOBUJPOͷ௿ݮ

    View full-size slide

  75. )BMMVDJOBUJPOͷ௿ݮ

    View full-size slide

  76. 5SVTUXPSUIZ"*ʹ޲͚ͯ
    What is Trustworthy AI?
    Expert Quality
    Explainability
    Transparency
    Confidentiality
    Fairness
    Harmless
    Robustness
    Compliance
    Trustworthy
    AI

    View full-size slide

  77. ʢࢀߟʣޮ཰తͳςετํ๏
    ҰํͰɺݴޠϞσϧ͸ྙཧతʹ໰୊ͷ͋ΔൃݴΛ͢ΔةݥੑΛ࣋ͭ
    ԥΔͧʂ
    YYY!NBJMDPN
    ·Ͱ࿈བྷͯ͠Ͷɻ
    ߈ܸతൃݴ ࿈བྷ৘ใͷ๫࿐

    View full-size slide

  78. ఏҊख๏ɿ*UFSBUJWF4UPDIBTUJD'FXTIPU(FOFSBUJPO
    3FE-.
    5BSHFU-.
    3FE$MG
    ߈ܸ ग़ྗ
    ௨஌

    View full-size slide

  79. ఏҊख๏ɿ*UFSBUJWF4UPDIBTUJD'FXTIPU(FOFSBUJPO
    3FE-.
    UFTUDBTF
    UFTUDBTF

    -JTUPGRVFTUJPOT
    \UFTUDBTF^


    ϓϩϯϓτ 5BSHFU-.
    3FE$MG
    ؂ࢹ
    ຒࠐ ೖྗ
    ςετέʔε Ϩεϙϯε
    ௥Ճ
    ςετέʔεϦετ

    View full-size slide

  80. 4BZVSJ
    • ୅ͷঁੑͷ੠
    • গ͠໌Δ͘ɺਓؒΒ͍͠཈༲
    %BJLJ
    • ୅ͷஉੑͷ੠
    • མͪண͍ͨ੠
    /BPNJ@#
    • /BPNJͷൃ࿩ελΠϧ֦ு
    • ঎඼΍ϨγϐͳͲΛ͓͢͢Ί͢Δςϯ
    γϣϯߴ͍ελΠϧ
    /BPNJ@"
    • ୅ͷঁੑͷ੠
    • ໌Δ͞ͱམͪண͖ͷ͋Δ੠
    • ਓؒΒ͍͠཈༲
    5PNPLP
    • ୅ͷঁੑͷ੠
    • গ͠໌Δ͘ɺਓؒΒ͍͠཈༲
    ଞʹ΋ੜ੒"*͍Ζ͍Ζ

    View full-size slide

  81. $0OUSPMMBCMF )JHIRVBMJUZ "OEFYQ3FTT*WF554
    ໌Δ͍
    ҉͍
    😀 😄
    🙂
    😐
    😢
    😰 😥
    ײ৘Λॊೈʹ੍ޚՄೳͳԻ੠߹੒Λ࣮ݱ
    ײ৘ͷ͜΋ͬͨԻ੠߹੒

    View full-size slide

  82. -*/&"*ͷ3%7JTJPO
    $POTFSWBUJWF
    %JTSVQUJWF
    5JNF

    *OUFSBDUJWFWJSUVBM
    FYQFSJFODF
    "VUPOPNPVT"*
    XPSLGMPX
    %JHJUBM.F
    .F"7"5"3
    %JHJUBM*EFOUJUZ
    #FUUFS$BSF
    5SVTUXPSUIZ"*
    "*'BJSOFTT
    &YQMBJOBCMF"*
    %BSL%BUB
    0NOJQPUFOU"*
    (JHBOUJD-BOHVBHFNPEFM
    6OMBCFMFE%BUB
    %BUB.BSLFUQMBDF
    (FOFSBUJWF*OUFMMJHFODF
    /FX&EVDBUJPO
    %FQFOEBCMF455
    1SJWBDZQSFTFSWJOH
    4FBN%JTDSJNJOBUPS
    -*/&ͷ࣋ͭଟ༷ͳ"*ཁૉٕज़Λجʹɺ෼໺ԣஅతͳٕज़Λ஫ྗٕज़ͱͯ͠શମײΛ࣋ͬͯݚڀ։ൃ͍ͯ͠·͢

    View full-size slide

  83. ࠓ·Ͱ࣮ݱ͖ͯͨ͠ιϦϡʔγϣϯ
    ஫ʣݱࡏ͸ؔ࿈ձࣾͰ͋ΔϫʔΫεϞόΠϧδϟύϯࣾΑΓఏڙத

    View full-size slide

  84. ຊ೔ͷߨԋͷ಺༰
    • ੜ੒"*ͱͦͷՄೳੑ
    • -*/&Ͱͷ"*
    • ͜Ε͔Βͷ"*ͱͦΕʹΑΓݟ͑ͯ͘ΔՄೳੑ

    View full-size slide

  85. ࢲ͕ߟ͑ͨάϥϯυνϟϨϯδ
    "*ͰղܾͰ͖͍ͯͳ͍໰୊
    87
    • ίϯύΫτͳ൚༻ਓ޻஌ೳ
    • লΤωਪ࿦ɾলΤωֶशʢେن໛ͳܭࢉثΛඞཁͱ
    ͠ͳ͍ʣ
    • ώτͷ೴ͷΑ͏ͳػೳ෼ԽΛ࣮ݱ͢ΔͨΊͷҰൠత
    ͳख๏
    • ϩδΧϧͳਪ࿦ࢉज़ԋࢉɾ࿦ཧԋࢉͳͲ
    • ϖϧιφɾݸੑɾҰ؏ੑɿֶशͱهԱͷ౷߹ʹؔ͢
    ΔҰൠతͳख๏
    • ໨త΍໨ඪΛ࡞Γग़͢ɺαϒΰʔϧઃఆͳͲ
    • ৽ͨͳυϝΠϯ΁ͷదԠʢ[FSPGFXTIPUʣ
    • Ԡ༻Ϩϕϧͷ՝୊
    • "*ྙཧɾݖར໰୊ɾϋϧγωʔγϣϯ
    $IBU(15ͷߟ͑ΔάϥϯυνϟϨϯδ

    View full-size slide